Designing an AI-assisted system for building, explaining, and fixing formulas
DEVELOPER TOOLS | ENTERPRISE B2B SAAS
PROJECT OVERVIEW
OVERVIEW
Formulas are powerful but notoriously hard to create, debug, and understand. This Unqork project explored how AI could act as a collaborative assistant inside a technical creation workflow, helping users build and fix formulas using natural language while preserving control and trust.
The result was an AI Formula Companion: a contextual popover that lives directly inside a formula cell and supports creation, explanation, and error recovery without forcing users to leave their work.
GOAL
Design an AI-powered experience that:
Reduces friction in formula creation and debugging
Keeps users anchored in their workflow
Adds guardrails without removing user agency
Scales from beginner to advanced users
SKILLS
AI Interaction Design
System & State Modeling
UX for Technical Workflows
Error Handling & Guardrails
ROLE
Lead Product Designer
TIMELINE
3 Weeks (September 2023)
PLATFORM
Web
PROBLEM
Users building formulas face three persistent challenges:
High cognitive load: Complex syntax and nested logic are difficult to reason about.
Error-prone workflows: Small mistakes lead to broken formulas and frustrating debugging cycles.
Fragmented help: Users jump between product docs, forums, and external references to solve a single issue.
At the same time, formulas represent structured intent, making them an ideal candidate for AI-assisted translation from natural language to logic.
DESIGN RESEARCH
Remixing autocomplete, AI<>human interactions, and Excel
Luckily, I did not start the design process from scratch. This project built on a prior Formula Autocomplete initiative I designed in close collaboration with platform engineers:
That work clarified key jobs-to-be-done:
Discover available formulas and arguments
Avoid syntax errors early
Learn what a formula does without leaving the builder experience
The AI companion layered intelligence on top of this foundation, rather than replacing existing tooling.
I also knew I wanted to combine current AI<>user interactions with traditional formula building helpers. I studied current tools including:
Generative AI: Google Duet AI, Notion AI
Formula building: Excel, Google Sheets, and Coda
Key insight: The most effective AI tools are contextual, assistive, and predictable—not separate chat destinations.
DISCOVERY
Unpacking creators’ formula writing mental models
Again, because I had led a prior Formula Autocomplete initiative, I had existing research about how humans approach the formula building process. This entailed qualitative interviews with both internal Unqork users (Solutions Architects, App Specialists, Product Managers) and external participants to understand how they build formulas in several applications including Excel, Tableau, and Unqork.
Key insights:
Users consult multiple sources to resolve a single formula issue
Syntax errors are the most common and frustrating failure mode
Visual structure (color, spacing, line weight) improves comprehension
Complexity stems from nesting, not advanced functions
Users want to save broken formulas and return later
DESIGN PROCESS
Building a Reliable Human<>AI Workflow
Interaction Design
I introduced a contextual formula popover, which is a user interface element that typically displays complex calculations, data previews, or formula details in a small, temporary overlay (popover) linked to a specific formula. Additionally, because the Unqork platform did not have an existing chatbot UI, I created new interaction patterns: recommended responses, user inputs, and AI responses. Each state was visually distinct to reinforce clarity, trust, and scanability.
Defining the System
I mapped the system states the AI needed to support:
Open Formula Builder
Create a formula using natural language
Explain an existing formula
Modify a working formula safely
Diagnose and fix broken formulas
Create a formula using natural language
Get help changing an existing formula
Fix a broken formula in seconds
Iteration & Constraints
Early designs were refined through close collaboration with UI and platform engineers. Based on technical constraints and research findings, I:
Adjusted component behavior to align with system limitations
Ensured AI suggestions were non-destructive and reversible
Designed error-handling states that preserved user control
The focus shifted from novelty to predictability and trust, which are critical in technical AI-assisted tools.
DEMO
Say goodbye to complicated formulas
I partnered with a platform engineer through implementation, providing detailed interaction specs and reviewing builds for UX fidelity.
The proof of concept was showcased at AWS re:Invent 2023, demonstrating:
A working AI-assisted formula experience
A scalable interaction model for AI-native creation tools